acceptodds
Under review as a conference paper at ICLR 2027

Few Step Variable Length Sampling with Branching Flow Maps

Abstract

Few step "flow maps" target the same problem space as diffusion and flow matching, but at vastly lower inference cost, which is becoming increasingly important with the large-scale use of inference-time reward alignment strategies. However, most standard diffusion, flow matching, and flow map models assume the length of the output is known a priori, which limits their usefulness, especially in scientific domains such as conditional small molecule or protein design. While variable-length flow matching approaches exist, they are inefficient at inference time, requiring many neural function evaluations. Few-step flow maps operate on continuous spaces, and it is unclear how to readily extend these to length variation, which is structurally a discrete property. Here we describe Branching Flow Maps, which enables few-step variable length generation on both discrete and continuous state spaces via element splitting and deletion. We demonstrate that our construction can compose with various established flow map approaches, affording efficient variable length generation with dramatically fewer neural function evaluations. We characterize the behavior of Branching Flow Maps on variable-length unconditional and conditional sampling tasks in synthetic and scientific domains, including small molecules and proteins, and we demonstrate accelerated reward-tilted sampling, with a sampling process that is now able to explore length as an additional dimension.

Primary area:
generative models
Keywords:
Flow Maps, Variable Length, Generative Models

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_Upw1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes Branching Flow Maps (BFMs), a few-step generative method for variable-length data. Time is divided into fixed stages, with continuous fixed-dimensional flow dynamics within each stage. Each element is additionally augmented with a continuous action variable that determines, at stage boundaries, whether the element is kept, deleted, or split, thereby allowing the dimensionality of the process to change. The method is trained using branching-forest conditional paths and finite flow-map distillation. Experiments on synthetic data, QM9 molecules, and antibody generation demonstrate effective variable-length generation with few neural function evaluations.

Critical strengths

The forest-based branching construction is intuitive and well motivated, providing a natural mechanism for variable-length generation through coordinated splitting and deletion events. The experiments also demonstrate the efficiency of BFM, achieving competitive generation quality with substantially fewer neural function evaluations, along with additional runtime improvements in antibody generation.

Critical weaknesses
  1. The efficiency evaluation relies largely on NFE, which does not account for the computational cost of each evaluation. In BFM, the population size changes across stages, and the flow model evolves continuous action variables together with the state, so the cost per NFE may vary across methods and stages. Although wall-clock runtime is reported for the antibody experiment, similar wall-clock or FLOPs-based comparisons should be provided for all experiments to more convincingly support the claimed efficiency gains and their generalizability across architectures and applications.

  2. The number and locations of stage boundaries are pre-specified rather than learned. Since the paper shows that the number of stages can affect few-step performance, it would be helpful to provide more intuition or empirical analysis on how to choose these stages for a new data distribution and how sensitive the method is to this hyperparameter.

Questions for authors
  1. Can you provide intuition/analysis on how to choose the number and locations of stage boundaries, which is transferable to other data?

  2. Can you provide wall-clock or FLOPs-based comparisons for all experiments?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Upw1

We thank the reviewer for the review. While it has taken us some time to do so, we believe we have addressed all of the reviewer's concerns, and we hope that the reviewer can consider raising their score in the reviewer/AC discussion period. We address the specific points below:

W1 / Q2: "Computational cost"

“Efficiency evaluation relies largely on NFE… can you provide wall-clock or FLOPs-based comparisons”

An excellent point that was a common concern among reviewers. We now have architecture and compute matched comparisons (see response to reviewer qFy3, W2 / Q2, for more detail), and show NFE and wall clock results (Figures 7–8 and 23; Tables 1, 9, 13 and 16). In summary, with matched architectures there is no substantial overhead for Branching Flow Maps relative to many step approaches (e.g. Branching Flows) and the NFE results closely track the wall-clock results (except when many-step methods collapse into degenerate very short samples at low NFE). We do not provide wallclock sampling timing for the "toy" sinusoid model (just as one would not typically provide these for the classical "spiral" example used in flow matching), but we do include training times.

W2 / Q1: “Intuition / analysis on how to choose the number and locations of stage boundaries”

This is an excellent point and we thank the reviewer for bringing it up. We now have experiments that, for QM9 and Sinusoid tasks, directly compare 2, 3 and 4 stages (Figures 9–10), and then for the standard 4 stage setup we compare linear interval boundaries vs boundaries that are shifted earlier and vs boundaries that are shifted later (Figures 11–12). Note that 2 is the minimum stage count for a Branching Flow Map model. For stage count, the behavior is mostly similar, where the main difference is that stage count puts a lower bound on the NFE you can attempt, and lower stage counts slightly improve the performance at lower NFE. For stage boundary location, we see a mild consistent benefit of "sqrt" spacing (0, 0.5, 0.707, 0.866, 1) for both QM9 and sinusoid (Section 4.2; Appendix D.2; Tables 4 and 10; Appendix E.6).

Our intuition here is that this choice is much like the choice of "schedule" in standard diffusion or flow matching, and we expect there to not be a universal "best" answer here for all tasks, and it will depend, for example, on the length distribution of what you're trying to learn, as well as interacting with other design choices, such as the starting length distribution. For example, for tasks where you have an unknown but relatively tight length distribution, it might suffice to initialize the number of elements close to the mean of that distribution, and then have only a single interval boundary where the length can change, for example, but in other tasks allowing multiple branching stages might be beneficial.

We now discuss this, and describe these experiments, in the manuscript (Sections 4.2 and 5; Appendices D.2 and E.6).

Reviewer_F176OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents Many-step Branching Flows, an approach for sampling using few-steps in a space with variable length. This model is motivated by application such as molecule generation, where typically we can start with n atoms and evolve them via flow matching keeping the number of n atoms fixed, or alternatively permit branching, deleting and variable amount of atoms changing with the evolution of the system. Furthermore, another interesting development has been the many-steps samplers, that essentially circumvent the step-by-step iterative generation of diffusion models or flow models, instead skipping multiple steps in one go. Such variable length combined with many-step sampling scenarios although better suited for many tasks, bring a set of challenges related to how to efficiently constructed such models and their properties.

This paper proposes a particular methodology based on combining a continuous branching action that is interpreted into actions that keep, delete, or split decisions, which can be a source of correlated branching and utilizing latent branching forests during training. The empirical findings of the paper are good and explore diverse set of scenarios (synthetic alternating-sinusoid task, QM9 molecule generation, and conditional antibody sequence completion).

The results indicate improvements in sample quality at small numbers of neural function evaluations compared with direct low-step integration of instantaneous models. The antibody experiments are particularly interesting because the model can jointly infer the lengths and contents of multiple missing regions.

Critical strengths
  • A conceptually clean construction. The separation between continuous within-stage evolution and discrete boundary operations is a natural way of modelling the many-steps-then-branch.
  • Substantial few-step empirical performance improvements.
  • Broad empirical coverage.
Critical weaknesses
  • The main theoretical result establishes idealized process correctness, not correctness of learned few-step maps
  • The language in the introduction is sometimes convoluted and was hard for me to follow. A more direct and clear writing there is needed.
  • Neural function evaluations do not provide a fully compute-matched comparison, especially when comparing models with orders of magnitude of difference in number of parameters.
Questions for authors
  1. Can the authors provide explicit sufficient conditions under which the within-stage marginal ODE transports the initial distribution to the boundary distribution? In particular, can Lemma C.2 be established directly for the proposed interpolation?

  2. Is there a useful bound connecting finite-flow-map prediction errors to errors in the decoded population distribution?

  3. Can the authors show that the learned action variables capture the required correlations?

1 comment
AuthorsOpenReview ↗

Response to reviewer F176

We thank the reviewer for their review, and appreciate their assessment of the construction as conceptually clean. While it has taken us some time to do so, we believe we have addressed all of the reviewer's concerns, and we hope that the reviewer can consider raising their score in the reviewer/AC discussion period. We address specific points below:

W1: “Main theoretical result established idealized correctness, not correctness of learned few-step maps”

The reviewer is correct to point this out, but we note that the standard way to train few step models (see https://arxiv.org/pdf/2505.18825, App. F.4) uses stop gradients, which introduces theoretical intricacies. The stop-gradient objectives share a fixed point with the “true” objectives but do not yield the same backward pass update. While there is work on the theoretical soundness of these or alternative objectives (https://arxiv.org/pdf/2609.16222),(https://openreview.net/forum?id=sWVDqeX4ux) we view proofs of this as being outside the scope of our paper.

W2: “Language in introduction convoluted and hard to follow”

Upon rereading with fresh eyes, we agree with you. We have rewritten the introduction (Section 1), and we hope you will agree that it is substantially clearer and easier to follow. We thank you for raising this point.

W3: “Neural function evaluations do not provide a fully compute-matched comparison”

This is an excellent point, also brought up by other reviewers, which we have now thoroughly addressed. Please see response to reviewer qFy3 (W2 / Q2). In summary: we now have architecture and compute-matched comparisons for all three tasks, now allowing clearer attribution of our few-step sampling advantage over other variable length modeling frameworks (Appendices D.1, E.4–E.7 and F.3–F.6; Tables 2–3, 11 and 14–15).

To the specific questions:

Q1

“...provide explicit sufficient conditions under which the within-stage marginal ODE transports the initial distribution to the boundary distribution? … can Lemma C.2 be established directly for the proposed interpolation?”

We thank the reviewer for pointing this out, we agree it is better to introduce the conditions. We have now introduced the assumptions required for Lemma C.2 and give an explicit proof (Appendix C.2). The assumptions required for Lemma C.2 to hold are the same as the standard assumptions required for fixed-length ODE flow matching.

Q2

“Is there a useful bound connecting finite-flow-map prediction errors to errors in the decoded population distribution?”

This was a thought-provoking question to answer. Distributional bounds in the variable-length case are substantially more theoretically intricate than their fixed-length counterparts. In our setting the Wasserstein distance is hard to interpret since it requires a metric on the underlying space, and several different metrics all induce the disjoint union topology, which all yield different interpretations on the Wasserstein distance. We chose a metric that assigns distance d/(1+d) to errors within lengths and distance 1 to length errors, which has the rough interpretation “Within-length errors saturate below one, and all cross-length errors are considered equally bad and worse than any within-length error”.

Since we use several different few-step distillation methods and since bounds for few-step models are still largely an immature field we elected to prove a general result stating that, if one can bound the within-length Euclidean Wasserstein distance when starting from the correct source distribution one can translate this to a bound on the variable-length Wasserstein distance based on our metric (see Appendix C.3, Theorem C.7 and Remark C.8), allowing us to inherit any bounds for fixed-length models that may arise or already exist in the literature.

Q3

“Can the authors show that the learned action variables capture the required correlations?”

In theory, from the perspective of the within-stage flow map, the actions are just another component of the state vector, and any flow map that can learn correlations between discrete elements should similarly learn the required correlations for the action variables.

In practice, for real examples these correlations will be extremely difficult to characterize. What we can show, however, is that sampling quality degrades severely when we break the correlations between the action and the state, or between the actions at different positions.

We now include an appendix section (Appendix E.8, Table 12) that shows that breaking action/state correlations reduces BFM 8 NFE QM9 connected validity from 86% to 54%, and sampling per-position marginal actions reduces connected validity to 10%. This shows that, to achieve the sampling performance reached by BFM, we must to some extent be learning rich action and action/state correlations.

Reviewer_qFy3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Branching Flow Maps (BFM) for few-step variable-length generation in continuous, discrete, and mixed state spaces. The central idea is to augment each element with a continuous action variable encoding its offspring count, evolve these actions jointly with the element states, and execute splitting, retention, or deletion at predefined stage boundaries. This construction enables the use of fixed-dimensional flow-map methods within each stage while modeling dependencies among structural decisions. The authors evaluate the approach on a synthetic task, QM9 molecule generation, and conditional antibody sequence completion, and demonstrate its use in reward-guided search.

Critical strengths

1.The interface between continuous transport and structural changes is technically valuable. Encoding offspring counts as continuous actions and jointly transporting them with the states provides a principled way to coordinate discrete structural decisions in few-step generation. The staged construction also offers a useful interface for adapting existing flow-map methods to variable-length settings.

2.The evaluation covers complementary tasks and state spaces. The synthetic task examines dependencies among sequence length, geometry, and discrete labels; QM9 tests joint generation of molecular coordinates and atom identities; and antibody completion evaluates conditional generation with unknown region lengths. Together, these experiments provide useful evidence that the construction applies across different generation settings.

Critical weaknesses

1.Limited evidence of conditional compatibility. The antibody evaluation primarily measures agreement with pooled natural reference distributions, including sequence length, 3-mer-based diversity, CDR3 length, and AbNatiV scores. While these metrics demonstrate distributional recovery, they do not directly establish whether the generated regions are compatible with the particular retained sequence context. More direct evidence is needed to assess whether the model meaningfully uses the supplied context.

2.Model capacity and training differences confound the comparison. BFM contains 254M parameters, compared with approximately 95M for Branching Flows and Edit Flows, and undergoes additional finite-map training and polishing. Consequently, its improvements at matched NFE may partly reflect greater model capacity and additional optimization. A capacity-matched comparison, together with explicit training-budget information, would help isolate the contribution of the proposed methodology.

Questions for authors

1.The antibody metrics primarily assess agreement with pooled natural reference distributions. Could the authors provide more direct evidence that the generated regions are compatible with the particular retained sequence context?

2.BFM contains 254M parameters, compared with approximately 95M for Branching Flows and Edit Flows, and receives additional finite-map training and polishing. To what extent do these differences contribute to the improvements at matched NFE? Could the authors provide a capacity-matched comparison and report the total training compute?

1 comment
AuthorsOpenReview ↗

Response to Reviewer qFy3

We thank the reviewer for their comments. While it has taken us some time to do so, we believe we have addressed all of the reviewer's concerns, and we hope that the reviewer can consider raising their score in the reviewer/AC discussion period. To the specific points raised:

W1 / Q1: "Conditional compatibility"

“Could the authors provide more direct evidence that the generated regions are compatible with the particular retained sequence context?”

In addition to our metrics on whole sequences, we now include a specific demonstration of conditional learning in our antibody model, where we select one complementarity determining region (CDR) of the antibody sequence and examine another CDR, both in the empirical data (hold-out sequences) and from our Branching Flow Maps sampler. We include a visualization of this (Figure 22 in our revised manuscript; Section 4.4 and Appendix F.6) showing that, for example, CDR2 is strongly dependent on CDR1, in a way that matches the empirical data distribution. While for sequence learning it is generally difficult to evaluate conditional distribution matching (since conditioning on anything too specific means we won't have enough samples to characterize the conditional distribution), this shows that the generated sequences are strongly conditionally dependent on the supplied context.

W2 / Q2: “Capacity and training differences”

This is an excellent point, which we now address in the manuscript. For antibody models: for a size-matched comparison, we do two things:

  1. We scale up the Branching Flows and Edit Flows models to match the size of the Branching Flow Map model.
  2. Where previously the few-step BFM distillation involved expanding the model size, we now employ a few-step distillation strategy that fine-tunes the model at exactly the same size.

This means that Branching Flows, Edit Flows, BFM "instantaneous" and BFM few-step are all matched in size (Appendices F.3 and F.6). In terms of training budget, we consider Branching Flows and Edit Flows models that match both the "instantaneous" BFM model, and extended training versions of EF and BF that match the few-step BFM distilled models in total wall clock training time (i.e. including both instantaneous and few-step training costs) (Appendices F.5–F.6; Tables 14–15). Results show that the difference between many-step EF/BF vs few-step BFM is not due to mismatch in model sizes. We also include sampling results for both NFE and wall-clock timing (Figure 23; Tables 1, 13 and 16).

For QM9 and sinusoid models we similarly include results from the few-step distillation strategy that allows size-matched models (Appendices E.4–E.7; Tables 9–11). We include wall clock timing for QM9 models as well (Figures 7–8; Table 9).

See Appendices F.3 and F.6 for architecture details and Figure 23 for sampling cost for the antibody models, and Appendices D.1 and E.4–E.7 and Tables 2–3 and 9–11 for the QM9 and sinusoid models.

Reviewer_vG9HOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose Branching Flow Maps, which enable few-step variable length generation on both discrete and continuous state spaces via element splitting and deletion. It is experimentally validated on several tasks including synthetic alternating-sinusoid task, QM9 molecules, antibody infilling, and reward-tilted sampling over atom span.

Critical strengths
  1. The method is well motivated as it explains why independent per-element event sampling breaks over large steps and thus the proposed branching approach with splitting and deletion is necessary for variable length generation.
  2. The method is evaluated on a multitude of diverse challenges.
Critical weaknesses
  1. Flow-map steps cannot cross a stage boundary, so the number of branching opportunities is fixed in advance and consumes the evaluation budget.
  2. The experiemental evaluation has some shortcomings:
  • The antibody comparison is not parameter-matched (254M for BFM vs ~95M for Branching Flows and Edit Flows), so low-budget gains partly reflect model size.
  • The main baseline, Expanding Flow Maps, is reimplemented from an unreleased method that the authors themselves say they could not faithfully reconstruct, and they substantively deviated from its description.
  • The reward-tilting experiment is a self-described toy on atom span with no chemistry-relevant objective, and at 160 amortized NFE per output it forfeits much of the few-step advantage.
Questions for authors
  • Please elaborate to what extent the fact that steps cannot cross stage boundaries can be a short-coming and whether you see ways to cope with these shortcomings?
  • Please elaborate to what extent the experimental comparisons are fair and rigorous (not different model sizes, not baselines that are inadequately reimplemented).
1 comment
AuthorsOpenReview ↗

Response to Reviewer vG9H

We thank the reviewer for their review. We address the specific points and questions below. While it has taken us some time to do so, we believe we have addressed all of the reviewer's concerns, and we hope that the reviewer can consider raising their score.

W1: “Flow-map steps cannot cross a stage boundary”

This is correct, and it does place a lower bound on the number of NFEs we can sample with. But in practice, where there are strong geometric constraints on valid samples (as is typical for molecules or proteins) few-step models often require a number of NFEs larger than the bound imposed by our construction. So while we do expect in some cases there might be a small cost to the additional "length sampling" capability, we do not envision this to always necessarily be a practical restriction. For example, the performance of Riemannian MeanFlow, when generating proteins, is still climbing at 10 NFEs (https://arxiv.org/abs/2602.07744).

Also note that it should be expected that sampling, for example, across one of four Branching Flow Map intervals is substantially less complex than the sampling across a single fixed-length interval from t=0 noise to t=1 data, since a single BFM interval must only learn a partial denoising.

We now clarify this point in the discussion (Section 5), and we thank the reviewer for highlighting it.

W2: Experimental evaluation

“...The antibody comparison is not parameter-matched ... so low-budget gains partly reflect model size.”

This is an excellent point, also raised by other reviewers. It is now fully addressed. Please see our response to reviewer qFy3 (W2 / Q2; Section 4.4, Appendices F.3 and F.5–F.6, and Tables 14–15).

“...Expanding Flow Maps, is reimplemented from an unreleased method that the authors themselves say they could not faithfully reconstruct”

Regarding Expanding Flow Maps, we were in a difficult position. Upon a first read of that manuscript, it seemed to aim to solve the same problem, which suggests that we should include it as a baseline. However, even after substantial efforts attempting to implement it (without available code as a reference) as we explain in the manuscript (Section 2.2.1; Appendices A and E.6–E.7) we are unable to demonstrate reasonable performance on our tasks. Further, as also explained (e.g. in Appendix A.1), we believe there are theoretical difficulties with the EFM construction. Their manuscript does not actually demonstrate variable length sampling in the continuous state setting: they sample "molecular conformers", whose size is predetermined. As of the time of this rebuttal, the EFM GitHub repository is still empty and so we are unable to refer to the code to resolve these difficulties.

We thus have two choices here: either A) decide that there isn't enough evidence to consider this as a baseline and exclude it entirely from our manuscript, in which case our novelty is strengthened because there is no competitor for Branching Flow Maps, or B) describe our attempt to adapt and fix Expanding Flow Maps for our tasks, where Branching Flow Maps shows a clear performance advantage (Figures 16–17; Tables 9–10). We chose option B since we prefer a fuller disclosure, but we would happily defer to the reviewer's final judgement if they instead suggest that we should take a more conservative path and remove our attempt at an implementation, and any comparisons, from our manuscript.

"The reward-tilting experiment is a self-described toy on atom span with no chemistry-relevant objective, and at 160 amortized NFE per output it forfeits much of the few-step advantage."

We thank the reviewer for bringing this up. We had originally selected "span" since it is such a visually clear objective, where the steering would be obvious to non-specialists (since our paper is about a general method, and not molecules specifically). But we agree that we should also include chemistry-relevant examples, and clarify that few-step models still give a substantial advantage in terms of NFE per output.

To address the "chemistry-relevant" point, in addition to the "span" example (which now also selects for validity and good molecule geometry), we include Topological Polar Surface Area, and cLogP (for lipophilicity) (Section 4.3; Appendix G.2). To address the "forfeits much of the few-step advantage", we compare few-step Branching Flow Maps to Branching Flows (which doesn't allow accurate few-step sampling) across a range of step increments and search budgets, showing that few-step variable length sampling allows the NFE budget to be spent on more aggressive search, since you spend far fewer steps per rollout, as expected (Appendix D.4, Figures 18–21; Appendix G.3).

We now clarify this point in the main text (Section 4.3), and include extensive reward skewing results in the appendix (see Appendix D.4, Figures 18–21).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision summary.

We thank the reviewers for their helpful comments. We have revised the manuscript and expanded the experiments to address the main points raised. We believe the additional experiments substantially strengthen the manuscript. The principal changes are:

  1. Retrained models and reran experiments to facilitate architecture- and compute-matched comparisons, with expanded reporting of training budgets and wall-clock sampling costs.

  2. Added theoretical analysis of error bounds in the Branching Flow Maps setting, connecting within-stage approximation errors to errors in the final variable-length distribution. We also made the assumptions and proof for within-stage transport explicit.

  3. Added systematic comparisons of stage count and stage boundary location, and expanded the discussion of the restrictions imposed by stage boundaries.

  4. Added evidence of conditional learning in the antibody model, comparing generated and empirical CDR distributions under different retained sequence contexts.

  5. Added experiments testing the importance of learned action dependencies, examining how sampling quality changes when action–state and inter-action dependencies are disrupted.

  6. Expanded the reward-guided sampling experiments to include chemistry-relevant objectives and comparisons across sampling and search budgets, demonstrating the benefit of few-step generation for inference-time search.

  7. Rewrote the introduction and reorganized the appendix to improve clarity. In light of the new models and experiments, we grouped related figures and tables to make comparisons easier to follow.

  8. Made various minor unrequested changes. For example, the Branching Flows QM9 models in the initial submission were switched from FP32 to BF16 part-way through training. We have retrained these models using BF16 throughout, matching the few-step model training setup, with no substantial difference to the results.

In the submitted PDF, we have attempted to track notable changes in purple text. While it cost us the entire discussion period, we elected to provide a revised version with all new experiments complete. If any reviewers have any further requests that they can trust us to make, we will of course include these in our camera ready version, should this be accepted.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.